Objects and scenes can be visually characterized by localized features and their appearance, plus the spatial relations between them. I present our recent work on learning hierarchical representations for such features based on statistics of feature cooccurrences. The learned features combine view-specific aspects within a single graphical model. Their hierarchical structure as well as bottom-up and top-down evidence integration using Nonparametric Belief Propagation loosely resemble that of higher biological visual systems. I will discuss examples of unsupervised learning for detection and recognition, and of feature learning during and for closed-loop interaction.